计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (12): 106-112.DOI: 10.3778/j.issn.1002-8331.2203-0043

• 模式识别与人工智能 • 上一篇    下一篇

简化且多层结合的知识图谱卷积网络推荐算法

杨兴耀,李想,于炯,郑捷,黄仲浩   

  1. 新疆大学 软件学院,乌鲁木齐 830008
  • 出版日期:2023-06-15 发布日期:2023-06-15

Light and Multilayer Combined Recommendation with Knowledge Graph Convolutional Network

YANG Xingyao, LI Xiang, YU Jiong, ZHENG Jie, HUANG Zhonghao   

  1. School of Software, Xinjiang University, Urumqi 830008, China
  • Online:2023-06-15 Published:2023-06-15

摘要: 将知识图谱和图卷积网络共同用于协同过滤可以很好地提升推荐质量。原图卷积网络在每层节点更新时使用的特征转换矩阵对推荐任务来说是冗余的,而且现有的基于知识图谱卷积网络的推荐方法使用最后一层卷积层学到的节点表示作为目标节点的最终向量表示,而忽视了前几层目标节点的向量表示。针对这两个问题,提出了简化且多层结合的知识图谱卷积网络推荐算法(LMCR)。该算法在各阶卷积层执行节点更新操作时,去除特征转换矩阵并根据数据的稀疏性有选择地保留或舍去自连接;在计算目标节点的最终表示时,结合各卷积层学到的目标节点表示。在MovieLens-20M和Last.FM数据集上进行的实验显示LMCR优于其他推荐模型,并验证了去除特征变换矩阵和多层结合操作对提升推荐性能的有效性。

关键词: 推荐系统, 协同过滤, 多层结合, 图卷积网络, 知识图谱

Abstract: Knowledge graph and graph convolutional network used them simultaneously in collaborative filtering can greatly improve the recommendation quality. The original graph convolutional network uses feature transformation matrix in node update, which is redundant for the recommendation model. And the existing based on knowledge graph convolutional network recommendation method uses the node representation of the last convolutional layer as the final vector representation of the target node, ignoring the target node representation of the previous several layers. To solve these two problems, a light and multilayer combined recommendation with knowledge graph convolutional network(LMCR) is proposed. The algorithm removes the feature transformation matrix and preserves or removes self-connection selectively according to the sparsity of the data when performing node update operation at each convolution layer. This paper combines target node representation learned at each convolutional layer when calculating the final representation of the target node. Experiments on Movielens-20M and Last. FM datasets show that LMCR is superior to other recommendation models, and verify the effectiveness of removing feature transformation matrix and multilayer combination operation in improving recommendation performance.

Key words: recommender system, collaborative filtering, multilayer combination, graph convolutional network, knowledge graph